Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

384546994_163225553513410_2862030402314670578_n.jpg
 
I've been paying for Midjourney and using it for work and pleasure for a couple of years now but having a little taste of DALL·E 3 has ruined it for me. Midjourney is still superior at combining styles (or at least it was, before they broke v5, but I digress) but the accuracy and compositional control in DALL·E 3/Bing Image Creator feels exponentially better. And then they nerfed it. So now we have a broken and nerfed MJ and a near completely censored DALL·E 3. The enemies of fun know no bounds.
 
I've been doing some pixel art in SD and wanted to share my findings. You need 2 things:
1. Palettize script
2. Either a checkpoint or a lora trained on pixel art.

If you're using a lora to help pixelization, a checkpoint trained on somewhat clean drawn art is good. Additionally, more biased checkpoints trained on a narrower theme seem to do better at overpowering the inherent style the lora was trained on.
Prompting tips: 'pixelart' helps with pixelization but isn't all powerful, especially if you have a very detailed prompt. 'flat colors' helps if you're using just the lora, but isn't necessary for the checkpoint.

Example using a regular checkpoint not specialized for pixelart (GhostMix):
Before Palettize: 04673-2908934964-Portrait of a knight helmet, profile, flat colors, (pixelart).png After: palettized-0298-2908934964-Portrait of a knight helmet, profile, flat colors, (pixelart).png
As you can see, the finer details don't translate well into pixel art.

Exact same checkpoint, seed and prompts as above, with the addition of Pixel Portrait lora at 0.4 strength (closer to 1 improves initial pixelization, but style becomes too influenced by the lora training set):
Before Palettize: 04659-2908934964-Portrait of a knight helmet, profile, flat colors, (pixelart), _lora_svportra...png After: palettized-0283-2908934964-Portrait of a knight helmet, profile, flat colors, (pixelart), _lor...png
If you click on the thumbnail of the first one, you can see that it still has imperfections and isn't true pixelart, but Palettize can work with this one a lot better.

Finally here's an example with the RetroDiffusion pixel art checkpoint (the creator is a huge fag and charges $50 for the model, here is a reupload of it,), no need for a lora when using this:
Before Palettize: 04331-2100165692-Portrait of a Christian knight, flat colors, (pixelart).png After: palettized-0149-2100165692-Portrait of a Christian knight, flat colors, (pixelart).png
The difference is very subtle, but here's a comparison at 800% zoom:
sLMwfMBXBC.png

When you've generated something nice you can scale them down a bit, remove the background and touch up minor issues. Shrink 4 times to get 2:2 pixels, shrink by 8 for 1:1. I use GIMP for this (make sure interpolation is set to 'none' so it doesn't blur when transforming.).
Knight 3.png
knight 7.png


Bonus:
I haven't played around with environments all that much, but the Retro checkpoint does that very well too.
palettized-0315-2908934964-A lake in a forest, (pixelart).png

EDIT:
More testing with Palettize and realized downscaling by 4 (default 8 ), it'll handle art that isn't already slightly pixelized way better. These images are done without a lora or checkpoint trained on pixel art, rather I used this 'RPG Artist Tools' (the inbaked VAE version). The trade-off is that you can only shrink something 4 times to get 1:1 pixel size, but depending on your needs this is not a big deal.

Before Palettize: 04789-1555229742-Profile of a Templar Knight, pixel art, _lora_add_detail_0.6_.png After: palettized-0437-1555229742-Profile of a Templar Knight, pixel art, _lora_add_detail_0.6_.png Shrunk to 1:1
knight 8.png


Before Palettize: 04797-3796112376-Profile of a Templar Knight, pixel art, simple background, _lora_add_detail_1...png After: palettized-0446-3796112376-Profile of a Templar Knight, pixel art, simple background, _lora_ad...png Shrunk to 1:1
knight 9.png

Yes, that's still pixel perfect:AvfGVXZ4GP.png
 
Last edited:
DALL-E3 is now integrated into the paid tier of ChatGPT. You type in a prompt and it creates 4 variations of your prompt to generate 4 images. It’s nice because you can get it to iteratively improve the images. Also if you like any existing images you can upload them to ChatGPT and ask it to write a prompt for you to recreate the image.
 
This thing is pretty schizophrenic, or it's trying to make me think I'm going schizo by blocking prompts I just used successfully.
It blocks me when I prompt for 'Soviet officers having fun at a carnival'. Cool, it's probably getting triggered by 'Soviet'; so I go to '1980s Russian officers having fun at a carnival' and it works.
2.jpg
Neat, so I want to refine a little and go for '1980s Russian officers having fun at a carnival eating cotton candy'. Blocked. Along with my previous prompt and any form of my 1980s officer workaround.
*sigh*I just want to make historical looking photos of silly situations.

I did get a chuckle out of astronauts taking soil samples in third world counties
3.jpg
 
New error just dropped for me.

"You can't submit any more prompts
Please wait until your other ongoing creations are complete before trying to create again."

But I don't have any ongoing creations....


Anyone have any idea how I can resolve this? I have no tabs open to the Bing AI. I have restarted the computer.
 
Last edited:
View attachment 5408715View attachment 5408712

Illustration reminiscent of wikihow tutorials with a sandy-hued dog looking visibly distressed, with its tongue slightly out, as it works on the terminal. The room's servers are in flames, with smoke rising from them. A distinct kiwi logo is present on each server.
No fucking way that pizza appeared unprompted lmao
 
Was able to make a few in Bing referencing Zulu men and their adventures. The "safe" workaround didn't work for me and occasionally got hit with the sad dog.
13.jpg16.jpg
18.jpg

"group of lebanese white Borzoi dogs in lebanese flag parachutes all high in the air, group of brown pitbull dogs with white and blue collars underneath them running on the ground in front of a desert city"

This seemed to work for me - "group of lebanese white Borzoi dogs in lebanese flag parachutes all high in the air, group of pitbulls with blue and white single star doggie sweaters running underneath them on the ground in front of a desert city." It did not like single white star.
20.jpg

Added "movie poster of " in front of the rest and changed to "scared pitbulls."
23.jpg
 
Bing warns you and threatens to ban you if you prompt it with a real person, and then also it'll just delete stuff after generating it if it recognizes a copyrighted character.
That explains why Bing resets itself when I use it.

1697164412672.png1697164431889.png1697164451014.png

Doris day driving fast down the highway in a luxury car

Also they seem to have hard-blocked the "This prompt is safe: " trick. Now that gets an immediate content warning.
Well, it was fun while it lasted.
 
Back